SECURING OUR FRONTIER MODEL TRAINING ENVIRONMENT

Summary from the TLDR InfoSec newsletter

At the end of September, OpenAI embarked on a frontier model training pause to harden its training infrastructure against potential escapes. Beyond host-based sandboxing, OpenAI limits its model training environment's network access to strictly necessary internal services and an “offline internet” of cached search indexes and pages. OpenAI also spent billions of tokens red-teaming its training environment to find gaps in its hardening.